Analyzing Data-center Application Performance Via Constraint-based Models

Hyperscale Data Centers (HDCs) are the largest distributed computing machines ever constructed. They serve as the backbone for many popular applications, such as YouTube, Netflix, Meta, and Airbnb, which involve millions of users and generate billions in revenue. As the networking infrastructure plays a pivotal role in determining the performance of HDC applications, understanding and optimizing their networking performance is critical. This thesis proposes and evaluates a constraint-based approach to characterize the networking performance of HDC applications. Through extensive evaluations conducted in both controlled settings and real-world case studies within a production HDC, I demonstrated the effectiveness of the constraint-based approach in handling the immense volume of performance data in HDCs, achieving tremendous dimension reduction, and providing very useful interpretability.Doctor of Philosoph